README.md

metadata

title: Backyard Arcade
emoji: 🕹️
colorFrom: yellow
colorTo: green
sdk: gradio
sdk_version: 6.16.0
python_version: 3.11
app_file: app.py
pypy: false
pinned: false
license: mit
tags:
  - build-small-hackathon
  - thousand-token-wood
  - track-2
  - gradio
  - smolagents
  - agent
  - p5js
  - off-brand
  - best-agent
  - best-demo
  - field-notes
  - track:wood
  - sponsor:openai
  - achievement:offbrand
  - achievement:fieldnotes

Backyard Arcade

Tiny games grown from strange prompts.

A local-first arcade director for Build Small Hackathon Track 2.

Hackathon: Build Small - An Adventure in Thousand Token Wood

Live app: Backyard Arcade

Demo video: backyard_arcade_hf_feature_demo.mp4

Article / field notes: https://huggingface.co/blog/AnshulRaj/backyard-arcade-small-hackathon

Social post: https://www.linkedin.com/posts/anshul-raj-372740202_backyard-arcade-a-tiny-arcade-director-powered-activity-7471909124197511186-_vaA

Hosted model: Qwen/Qwen2.5-Coder-32B-Instruct

Users describe a tiny game in natural language. A small agent chooses one of four safe render tools:

• dodge
• collector
• jumper
• snake

For strange mode, the agent can blend two render tags into one playable game. The AI designs the cartridge: title, theme, goal, difficulty, controls, hazards, collectibles, and twist. The renderer owns the game loop, collisions, controls, and Gradio iframe safety.

Space Settings

Set these Space secrets/variables:

• HF_TOKEN: Hugging Face token for Inference API access
• MODEL_PROVIDER=hf
• HF_MODEL=Qwen/Qwen2.5-Coder-32B-Instruct

For local development, .env can use HUGGING_TOKEN instead of HF_TOKEN.

For local Ollama development:

• MODEL_PROVIDER=ollama
• OLLAMA_MODEL=qwen3.5:9b
• OLLAMA_HOST=http://127.0.0.1:11434